Session: 17-01-01: Research Posters

Paper Number: 150878

150878 - A Machine Learning Approach to Sensitivity Analysis: Maize Stalk Flexure and Strength 

INTRODUCTION: Maize (corn) is the most produced grain crop world-wide. However, approximately 5% of the maize harvest is lost each year due structural failure of the stalk. Maize stalk failure is governed by localized buckling failure (Brazier buckling), which is influenced by a complex interplay between geometry and material properties. In this study, a unique combination of experimental data, computational modeling, machine-learning, and uncertainty quantification methods were used to investigate this problem. The goal of this research was to identify the factors that have the strongest influence on maize stalk strength.

METHODS: First, machine learning techniques wereused to create a fully parameterized model of the maize stalk morphology. The parameterized model was based upon a database of 1000 micro-CT scans of maize stalks. The resulting model consisted of 96 geometric parameters and 12 physical material properties. The parameterization scheme allowed independent control of each physical feature of the stalk. The parameterized maize stalk model was rigorously validated using stochastic modeling to account for inherent uncertainties in material properties. A multi-dimensional copula was used to create synthetic maize stalk models that obey the same set of covariance patterns observed in the original data set.

Next, principal component sensitivity analysis was performed to (a) identify underlying patterns of variation within the maize stalk model; and (b) reduce the parameter space of the sensitivity analysis. The principal component sensitivity analysis was based upon a data set consisting of 900 specimen-specific parameterized maize stalk models. Principal component analysis captured and characterized important variation patterns that existed between the various parameters of each specimen-specific model. By controlling each principal component individually, the dimensionality of the parameter space was reduced dramatically.

Finally, a sensitivity analysis was performed on the material properties of the maize stalk. This was accomplished using a stratified Monte Carlo sampling approach with one-at-a-time variation in each material property.

RESULTS: The parameterized model accurately captured the shape of actual maize stalks. Validation studies revealed that the parameterized model was predictive of maize stalk stiffness and strength. The principal component sensitivity analysis revealed that variation in maize stalk strength is highly influenced by a single dominant pattern of variation and that the remaining principal components contributed relatively little to the overall variation in stalk strength. The dominant mode indicated by principal component analysis corresponded closely with theoretical predictions. Reduced-order models created using this approach were highly predictive of non-parameterized models.

A sensitivity analysis performed on material properties revealed that buckling instabilities were influenced by material properties to a greater extent than flexural stiffness. Finally, we demonstrated that this model can be used to create an unlimited number of synthetic stalks from within the parameter space.

CONCLUSIONS: Machine-learning methods provided an efficient and effective means for more completely understanding the factors that influence maize stalk strength and stiffness. Results indicate that a single dominant pattern of shape is the primary determinant of maize stalk strength. Results also indicate which material factors are most influential on maize stalk buckling. The resulting machine-learning model of maize stalks will enable the future implementation of optimization studies, and can be used to create computational models of maize stalks with any desired combination of geometric and material properties while still preserving important variation patterns that are observed in nature.

Presenting Author: Douglas Cook Brigham Young University

Presenting Author Biography: Dr. Cook earned a bachelor’s degree in mechanical engineering from Utah State along with minors in mathematics and Mandarin Chinese. He received Masters and PhD degrees in mechanical engineering from Purdue University. His research has been supported by a CAREER award from the US National Science Foundation, as well as grants from the US Department of Agriculture, the NSF, and industry. His research findings have been published fields as diverse as acoustics, biomechanics, biomedical engineering, agronomy, medicine, and botany.

Authors:

Joseph Carter Brigham Young University
Ryan Hall Brigham Young University
Douglas Cook Brigham Young University